LUCENE-8962: Merge small segments on commit - #1617
Conversation
…he presence of deletions
… work accordingly
IW might drop segments that are merged into a fully deleted segment on the floor and deletes the newly created files right away. We should not include these segments in a commit since we can't guarantee valid ref-counts on these files.
|
@mikemccand can you take a look |
mikemccand
left a comment
mikemccand
left a comment
There was a problem hiding this comment.
This looks really close! I left a couple questions.
I've been beasting this on 128 core Ryzen Threadripper ... it is up to 135 iterations and no failures, yay!
I have some small fixes to comments/javadocs/whitespace; I'll push them shortly after passing tests/precommit.
| } | ||
| } finally { | ||
| // ensure we assign this to close them in the case of an exception | ||
| this.mergeReaders = List.copyOf(readers); |
There was a problem hiding this comment.
Hmm is there some (sneaky) reason why we couldn't just do this.mergeReaders = new ArrayList<>(segments.size()); above and then append to that list? Instead of appending to private ArrayList and then making a copy in the end?
There was a problem hiding this comment.
I left a comment. I use immutable lists everywhere here it would be fatal if we modify this list outside of OneMerge. I think that justifies the copy.
| } | ||
|
|
||
| /** We only allow sorting on these types */ | ||
| private static final EnumSet<SortField.Type> ALLOWED_INDEX_SORT_TYPES = EnumSet.of(SortField.Type.STRING, |
There was a problem hiding this comment.
OK I see this was removed earlier in mainline for LUCENE-9330 -- it must've snuck back in as a merge conflict.
I'll re-remove now.
|
Wow! I didn't realize that capturing/cloning live docs at a point in time would be so complicated. I learned a lot from this PR about how deletes are handled. Thanks, Simon! |
| * Note: This settings has no effect unless {@link MergePolicy#findFullFlushMerges(MergeTrigger, SegmentInfos, MergePolicy.MergeContext)} | ||
| * has an implementation that actually returns merges which by default doesn't return any merges. | ||
| */ | ||
| public IndexWriterConfig setMaxCommitMergeWaitMillis(long maxCommitMergeWaitMillis) { |
There was a problem hiding this comment.
I am torn if we'd use a double or long here. I think we used double in other places but I do like a discrete value better...
@msfroh I didn't either :) it's been a while that I digged into merging code and I think the same is true for @mikemccand - I am impressed how far you got in this pretty complex class. I'd love to see you digging more and coming up with more stuff like this. Thank you for contributing this. I am always happy to help. This entire process from the first patch to the feature being integrated is a great example of how amazing OSS is. |
…e is doing; fix test case to use deterministic merging
|
merged into master. Thanks everybody involved!! |
Add IndexWriter merge-on-commit feature to selectively merge small segments on commit, subject to a configurable timeout, to improve search performance by reducing the number of small segments for searching. Co-authored-by: Michael Froh <msfroh@apache.org> Co-authored-by: Michael Sokolov <sokolov@falutin.net> Co-authored-by: Mike McCandless <mikemccand@apache.org>
4 participants
This PR revisits the merge-on-commit patch submitted by @msfroh a little while ago. The only change from that earlier PR is a fix for failures uncovered by TestIndexWriter.testRandomOperations, some whitespace cleanups, and a rebase on the current master branch. The problem was that updateSegmentInfosOnMergeFinish would incorrectly decRef a merged segments' files if that segment was modified by deletions (or updates) while it was being merged.
With this fix, I ran the failing test case several thousands of times with no failures, whereas before it would routinely fail after a few hundred test runs.
Co-authored-by: Michael Froh msfroh@apache.org
Co-authored-by: Michael Sokolov sokolov@falutin.net
This is the third try after the patch itself and #1552